Processing each single plane

Transcription

1 , Sept 30th, 2007 Performing a multiplane calibration on the dataset ptvsample The sample dataset is calibrated on a planar target imaged at three different heights; the procedure consists in the acquisition of the image coordinates of the target blobs at each level (referred as planes a, b, c) and the following composition of these data in a virtual 3d-target, on which the calibration procedure is performed. The working folder for the calibration is the subfolder cal, other paths indicated in the tutorial are children of cal. Processing each single plane Browse into cal and edit the content of the included batch file with the absolute paths of the ptv executable and the ptv.tcl file as installed on your machine. Launch the batch file. In the gui, choose Change P arameters Calibration P arameters and modify the fields as shown in figure 1. Figure 1: Parameters for plane a. l.delcastello@tue.nl 1

2 In cal, delete the file man ori.dat and copy man ori a.dat as man ori.dat. From the gui choose Calibration Show Calib. image and Detection, and check that all the dots listed in the target file (the central 9x9 dots for plane a) are properly detected and marked with blue crosses in each image. Proceed with Orientation with f ile, which uses the previously created man ori.dat file with the coordinates of four manually clicked preorientation points. Four guess.ori files with camera parameters are needed for preorientation. They consist in the position and orientation of the pinhole of each camera and the respective glass window (the refractive interface between air and glass) in the reference frame assigned through the coordinates of the calibration dots in the taget.txt files; the principle distance and the offset of the lens axis relative to the sensor centre. Precisely, the file is structured as: x cam y cam z cam pinhole position α cam β cam γ cam cam rotation vector R ij cam rotation matrix (not used in the guess files) x P y P lens offset f principle distance x glass y glass z glass interface position Once the four guess.ori files are manually edited, it is possible to visualise the reprojections of the grid dots on the images in order to check the quality of the guesses using the command Show initial guess: the yellow crosses should follow inside the images, if not their virtual image coordinates are anyway displayed in the terminal window. Now the preorientation can be executed through the command Sortgrid; check carefully that all the dots listed in the target file (9x9 for plane a) are properly assigned and marked with white crosses and point numbers, as in figure 2. Finally, Orientation does the real calibration: the results are printed in the terminal window and in new.ori files. If the algorithm does converge and the computed camera parameters are satisfactory (reasonably close to the measured positions used for the guess.ori files), the single plane calibration is checked. The image windows display the fitting error (enhanced for visibility) for each dot, and the gui reports the rms error for all the dots, see l.delcastello@tue.nl 2

3 Figure 2: Preorientation results for plane a. figure 3. Figure 3: Orientation results for plane a. Now some files with image and world coordinates of the dots have to be written, in order to perform the following multiplane combination. In Calibration P arameters, set to 1 the value of Calibrate with different z positions and execute once more Orientation. Check that.crd,.pix and fix files are created for each camera. Now repeat all the previous steps for the remaining calibration planes (b l.delcastello@tue.nl 3

4 and c in the example), paying attention to change the appropriate calibration parameters and the man ori.dat file; note that the points for manual preorientation are listed in the corresponding man ori.dat filenames. Combining the planes The last step of the procedure consists in performing the calibration on the virtual 3d-target. In Calibration P arameters change the appropriate fields for the plane m (as multiplane), and set both Calibrate with dif f erent z positions and Combine preprocessed planes to 1, as shown in figure 4. Figure 4: Parameters for plane m (multiplane combination). Execute once more Calibration Show Calib. image, and black images are shown. Finally Orientation performs the calibration on the virtual 3dtarget. As before, the results are shown in the terminal window and written to files: check that position and orientation of cameras and glass interfaces are consistent. Now it s time to improve the accuracy, adding one by one some distortion parameters, in order to take into account for lens offset and aberrations: go to Calibration P arameters, add one parameter, confirm with ok, run Orientation, proceed. Beside looking at the rms error in output, check l.delcastello@tue.nl 4

5 constantly how the output.ori files are modified: often the procedure tries to correct for wrong position and orientation with the use of k and p aberrations parameters, leading to fake results or not converging at all; it is thus essential that the procedure converges for all the cameras before introducing any lens distortion (as in figure 4). Further suggestions The software needs to be restarted before a new set of images is opened (e.g. for each plane) and whenever the orientation procedure does not converge or leads to inconsistent (wrong or NaNs) orientation values. In the latter case, it is always advisable to manually delete the latest.ori files and the raw.ori ones before launching the batch file. While optimising the calibration using the distortion parameters, it is convenient to backup the latest good.ori and.addpar files at each step. The content of consistent.ori files can be manually copied into the guess.ori files, in order to use it as a starting point for the following steps. Use the commands Checkpoints and AP figures for further checking the calibration results, especially while testing the influence of lens distortion parameters. In order to fully check the ability of the algorithm in 3d-positioning with the achieved calibration data, it is advisable to set the calibration image files as flow images (Main Parameters Name of image and Basename for sequence), restart the software and load the images with the Start command. Follow the three steps in the menu P retracking and finally press 3D coordinates to write an output file with the recovered world positions of the calibration dots. A shortcut to this is to perform a clean sequence of the commands: Show Calib. image, Detection, Orientation with f ile, Sortgrid, Orientation; now it is possible to use the calibration images already loaded following the menu P retracking and then 3D coordinates, without changing any parameter. In both cases the code writes in the folder res the file dt lsq with the recovered target coordinates in the first three columns, and displays in the command window the actual rms error for 3dpositioning. Two simple Matlab scripts are included in the sample working folder: display- Ori.m can be used to display the camera arrangement around the observation volume, useful when dealing with guess files and the first raw calibrations; displayrt is.m plots the recovered positions against the ones assigned in the target files. l.delcastello@tue.nl 5